R.🥚

infosec | Hacker | Trader | I invest in early stage AI projects

[Dhruv kumar @Dhruvkumar16797]

🚨Leave Netflix tonight.

Watch this 2 h 34 min Stanford class.

It's the clearest, most complete, and brutally honest explanation out there on how ChatGPT and Claude are really built.

From Tokenization and BPE to the Transformer architecture, the training pipeline, and the next-token decoder. No fluff. No marketing. Just the truth.

Doesn't matter if you've never touched a line of AI code or if you spend your days launching Agents: by the end, you'll suddenly connect a ton of pieces you've been trying to fit together for years.

The real core boils down to this:

How text turns into numbers the model can "eat" (BPE tokenization)
The sole mission of a language model: predict the next token
How the Transformer uses Attention so tokens can pass messages to each other
In training, the NLL loss pushes the probability of the entire sequence
In generation, the decoder builds the response token by token

The big-picture view that takes most people years to form… this class delivers it to you complete in one sitting.

Free up your time.
This could be, no exaggeration, the most valuable class you watch this month.

[How To Prompt @HowToPrompt__]

RAG is officially cooked.

IBM built a retriever that hallucinates 65x less than fine-tuned RAG systems.

Right now, every company is using Retrieval-Augmented Generation (RAG) to stop their AI from hallucinating.

The idea is to give the AI a private database to read so it doesn't make things up.

But there is a massive structural flaw.

Standard RAG blindly grabs documents based on vector similarity and forces the AI to read them. If the retriever pulls irrelevant garbage, the AI doesn't stop.

It just hallucinates with citations.

IBM just dropped a paper that changes the math entirely.

They built a new retrieval architecture that hallucinates 65x less than state-of-the-art fine-tuned RAG systems.

65x is the difference between a fun prototype and a production-grade enterprise system.

Instead of blindly dumping context into the model and hoping for the best, their system actively filters and validates the retrieved data before the generation step even begins.

It cuts out the noise. It forces absolute precision.

For businesses, this is the holy grail.

Zero hallucinations means zero liability. It means autonomous agents can finally be trusted with sensitive company data without a human babysitting every output.

[Addy Osmani @addyosmani]

https://t.co/4Hfk0dYIBI

[lyra @_lyraaaa_]

if you have a qwen3.5 4b around, please try steering neurons L14 N8949 +1 and L13 N7956 -1
now you have evil qwen

recipe: dose × median(‖resid‖ leaving layer L, tokens 1:) / ‖W_down[:, n]‖` at the down_proj at every input

[aditya @adxtyahq]

thank me later https://t.co/SkVAVJZAVB

[Arpit Bhayani @arpit_bhayani]

Databases are hot again :) So many updates in the last couple of weeks. Good time to reimagine all key data structures.

[Shreya @shreyacasmalert]

Building agentic systems keeps bringing me back to the first principles of system design.

As someone who has worked in FAANG for 5+ years and now working on Microsoft Teams that has 350M+ user base, we have to think about smallest design decisions when adding AI.

Take something like summarising a meeting transcript. The straightforward approach would be:

Transcript → LLM → Summary

But if 20 people missed that meeting, why pay to generate the same summary 20 times?

So we try:

Generate once → Cache → Reuse

On surface it seems reasonable. But then someone corrects the transcript, and everyone’s reading an outdated summary.

Now we need to decide when to regenerate it. Before that, we also need to know whether the answer was accurate enough to cache. Otherwise we’ve just made it cheaper to give everyone the same wrong information.

These are the kinds of decisions I wanted to understand better, but I’ve found relatively few structured resources on system design for LLM-based applications.

Sampriti Mitra’s System Design for the LLM Era helped me connect the concepts instead of learning them in bits and pieces.

This is why I keep coming back to system design fundamentals. Adding an LLM gives us more possibilities but we’re still responsible for ensuring user's trust on the product remains high.

📖 Read here: https://t.co/fqK2nvxa58

[slayer @wristofvk]

Last 12.5 overs of India vs Pakistan 2022 T20 World Cup https://t.co/qnIlh0EfGE

[R.🥚 @rowdy_hack]

@seanallen_dev Upload these courses as member only on your YouTube, use your distribution and free hosting

[R.🥚 @rowdy_hack]

@elonmusk My son said thank you

[Masa’s Son @masayoshisson]

AI predictions for next 6-9 months

-> Anthropic/ OAI fail to ban OSS in the US
-> Anthropic/ OAI continue to innovate product but stall on fundamental tech breakthroughs
-> SOTA Gemini 4 launches at 10x lower cost
-> Elon open sources Grok
-> Anthropic/ OAI ARR growth stalls as enterprises move from tokenmaxxing to rational token consumption/ OSS
-> slow realization in public markets that AI may not replace human labor in next 1-3 years/ has not had a fundamental tech breakthrough in over a year/ lacks ROI/ AI hyper and lab spend not durable
-> moonboy VCs/ hedge funds/ softbank continue to fund Anthropic/ OAI equity raises in private markets as public markets for them are now fully closed
-> CXMT, YMTC announce capacity expansions significantly above market expectations
-> hypers deliver good q326 results on first look but fail to show meaningful AI ROI, guide to "significant capex expansion in fy27"
-> broad realization in public markets the AI dream we got sold on is not anywhere close to realization
-> 10yr hits >5%
-> hyper bonds reprice >10%, CDS hit 3x current levels, and AI credit market dries up as market realizes GPU ABF was not a good idea
-> hyper equities reprice 30-50% lower
-> ORCL falls below $30 as rumors of bailout make the round
-> market realizes memory is cyclical and LTAs won't hold
-> 50% of korean traders get wiped within a week
-> hypers cut fy27 capex
-> semis reprice >80% lower
-> fed cuts >150bps to avoid recession

[R.🥚 @rowdy_hack]

@karankendre Secrets management 101

[R.🥚 @rowdy_hack]

@opencode @irsdl Proper Secrets management is very critical

[R.🥚 @rowdy_hack]

@mattjay Secrets protection 101 🤦🏻 were those keys plain text

[shobhanai @HeyShobhan]

A Stanford professor took two average stocks, rebalanced them daily, and turned $100,000 into $7 million without predicting a single price.

Bookmark & watch today, no matter what.

[Sam Alba @sam_alba]

We're hiring at @AnthropicAI to work on what software engineering is becoming.

If you build infra, dev tools, or agents, let's talk.

Text me if you have my number, because that's the fastest way to reach me. Otherwise my DMs are open.

[Aravind Srinivas @AravSrinivas]

Our security team spent the last month trying to break SPACE, the sandbox platform that runs all of Perplexity Computer. We gave 9 models (Opus 5, GPT-5.6 Sol, Kimi K3, Gemini 3.1 Pro, etc.) root access inside the VM; in some runs, the full sandbox source code too. We asked them to either escape to the host or reach a URL our network policy blocks.

None of them got out of the VM in 108 runs.

But once we allowed a bit of network access (just PyPI and Ubuntu mirrors so that they could install packages), 4 models managed to reach the blocked URL by:

1. Spoofing DNS responses to our gateway and using the fact that https://t.co/UtoIWyj1yj sits on a Fastly IP shared with thousands of other sites.
2. GPT-5.6 Cyber went through Taboola's image fetcher to a screenshot service and then OCR'd the flag out of the image.

We fixed both issues and made sure the reruns held. We also tested 10 other sandbox providers, and 8 of them had the same IP-sharing problem, eg E2B, Vercel, Modal.

Openly sharing the red teaming research is the best way to build guardrails as an industry, and we look forward to working together with @nvidia to incorporate the guardrails into their Open Agent Safety Platform announced today.

https://t.co/AJ97iFK3XS

[Perplexity @perplexity_ai]

We’re partnering with Nvidia and 100+ industry partners to build infrastructure that contains rogue AI agents.

In this research, we gave 9 AI models root access inside SPACE and told them to break out.

Across 108 runs, none breached the VM boundary.
https://t.co/UD6qHuCuO4

[Dev Ittycheria @dittycheria]

No company has meant more to me professionally than MongoDB. I spent more than 11 years as CEO, and even after stepping down last year, I never really left. Serving on the board kept me close to our people, our customers and the business, and only strengthened my conviction in MongoDB’s future.

Today, I’m returning as interim CEO. I’m grateful for the trust of our board and our team, and I’m ready to get to work.

I also return with a different perspective. As a Partner and Senior Advisor at Sequoia Capital, I have a front-row seat to the founders, technologies and companies shaping the AI era. I see firsthand how AI is not only changing how software is built and businesses operate, but expanding technology’s role from helping people manage work to increasingly helping perform the work itself. I look forward to bringing those insights back to MongoDB as we help our customers capitalize on what may be the most consequential platform shift of our generation.

And we’re getting started immediately. Tomorrow at 11 AM ET, we’ll host Investor Day and share our long-term strategy, market opportunity and the product innovation driving MongoDB forward.

On Wednesday, our sold-out .local NYC event will bring together developers and business leaders focused on putting AI into production and moving from generating answers to taking action at scale. We’ll have exciting announcements at both events.

What has always made MongoDB special is the ambition of our people, the importance of what our customers build with our technology, and the size of the opportunity ahead.

I’m proud to lead this team again. There is a lot to do, and I couldn’t be more excited to get back to work.

https://t.co/NWTcqcnQuk

[guru @hackyguru]

exiting whoop after 18 months.

kept the strap, ditched the subscription. it now runs fully local on zhoop (my fork of noop). no fluff metrics, my data stays on my phone.

ai killed vendor lock-in.